We started because the agents we loved kept failing the same way.
Every coding agent we used — Claude Code, Cursor, Aider, Continue — was brilliant at writing a function and amnesiac about the codebase it was writing into. They'd re-discover the same class hierarchy on every turn, burn tokens re-reading files they'd already read, and confidently propose changes that violated conventions buried three folders deep.
We tried every memory system we could find. Mem0 and Zep were built for chat transcripts, not code. Cursor Memories and CLAUDE.md are prose files injected into the prompt — great for preferences, useless for "enumerate every caller of DatabaseClient.connect." Letta calls the LLM to read from memory, which is exactly the cost we were trying to avoid.
None of them indexed code the way a compiler indexes code. So we built the one that does.
ArgosBrain is what happens when you point a language-server-aware graph at the "memory" problem and refuse to pay tokens on the read path.
The engine is called Neurogenesis internally — a Rust core that ingests your repo through SCIP, live LSP, or bespoke tree-sitter drivers depending on the language, builds a symbol graph + HNSW index, and serves it over MCP in under a millisecond. Zero LLM calls on the hot path. Works offline. Your code never leaves your machine.
We shipped it, published the benchmark, and made both open-source so that anyone — including our competitors — can verify the claims. That's still the deal.
